tun: windows: delay initial write

Otherwise we provoke Wintun 0.3.

tun/tun_windows.go

@@ -11,6 +11,7 @@ import (
 	"io"
 	"os"
 	"sync"
+	"sync/atomic"
 	"time"
 	"unsafe"
 
@@ -41,6 +42,7 @@ type NativeTun struct {
 	wt                        *wintun.Wintun
 	tunFileRead               *os.File
 	tunFileWrite              *os.File
+	haveRegisteredWriteBuffer int32
 	tunLock                   sync.Mutex
 	close                     bool
 	rdBuff                    *exchgBufRead
@@ -140,12 +142,7 @@ func (tun *NativeTun) openTUN() error {
 			}
 			return err
 		}
-		firstSize := (*uint32)(unsafe.Pointer(&tun.wrBuff.data[0]))
-		saved := *firstSize
-		*firstSize = 0
-		// Set the maximum buffer length with an invalid write.
-		tun.tunFileWrite.Write(tun.wrBuff.data[:])
-		*firstSize = saved
+		atomic.StoreInt32(&tun.haveRegisteredWriteBuffer, 0)
 	}
 	return nil
 }
@@ -324,6 +321,15 @@ func (tun *NativeTun) Flush() error {
 			return err
 		}
 
+		if atomic.CompareAndSwapInt32(&tun.haveRegisteredWriteBuffer, 0, 1) {
+			firstSize := (*uint32)(unsafe.Pointer(&tun.wrBuff.data[0]))
+			saved := *firstSize
+			*firstSize = 0
+			// Set the maximum buffer length with an invalid write.
+			tun.tunFileWrite.Write(tun.wrBuff.data[:])
+			*firstSize = saved
+		}
+
 		for {
 			_, err = file.Write(tun.wrBuff.data[:tun.wrBuff.offset])
 			if err != nil {